[systemd-devel] [PATCH v3] Added support for Uplink Failure Detection using BindCarrier

[Zbigniew Jędrzejewski-Szmek]

On Fri, Feb 13, 2015 at 09:53:28AM +0300, Andrei Borzenkov wrote:
> В Thu, 12 Feb 2015 15:25:08 +0100
> Zbigniew Jędrzejewski-Szmek <zbyszek at in.waw.pl> пишет:
> 
> > On Thu, Feb 12, 2015 at 12:52:00PM +0000, Rauta, Alin wrote:
> > > > Yes, but the updates need to be done for all links and I'm not sure adding this is a good thing.
> > > > I'm now having 64 links on the switch and I need the failure detection in networkd to be quite fast because however even now it's probably slower due to evaluating dynamically the BindCarrier strings when comparing this with the previous solution with an UFD group monitoring some interfaces and with some internal counters knowing exactly when to issue "link_down" for an interface. So adding "bound_by" and "bound_to" makes the solution even slower.
> > > 
> > > > How many times per second will you be avaluating this?
> > > Each time an event happens: a link appears, disappears, changes flags or names. 
> > Yes, I know the causes. I'm asking how often they can realisticly occur.
> > 
> 
> You misunderstand. It is not how often they occur but how fast reaction
> is. When (final) uplink goes down we want to bring dependent interfaces
> down as soon as possible.
Well... I don't think any string processing inside of networkd is going to
have a measurable impact. Things like process scheduling are always going to
trump that.

Not that I'm opposed to making the implementation computationally
faster. Just that reaction speed is a bad justification. The code should
be simple and clear, this is more important.

> > > > Besides this, having only one function "sd_network_link_get_carrier_bound_to" makes also sense because only the behavior of "bond_to" links is controlled by this feature. "bound_by" means almost nothing for an interface. A tool like "networkctl" may take into account to display only the "bound_to" links because that's what's relevant. The fact that "networkctl" displays both "bound_to" and "bound_by" it's a good thing, but it doesn't mean each tool should do that.
> > > 
> > > > If a link goes down, isn't the "bound_by" list useful to look at links which need to be checked and potentiallly brought down?
> > > It can be useful, that's why "networkctl" has the updates, but are talking about the showing functionality or about the run-time "up-down" game between interfaces ?
> > 
> > The latter.

Zbyszek